Sahm Rule Explained: When Does the Recession Alarm Trip?

US unemployment ticks up and a headline announces that the Sahm Rule has triggered. A reader who simply subtracts last month’s rate from this month’s will often get the wrong answer. The indicator compares smoothed rates and looks back for a low that can move as old months leave the window.

The Sahm Rule flags a rise of at least 0.50 percentage point in the three-month average US unemployment rate relative to its low over the preceding twelve months. It is a labor-market indicator. An official recession chronology considers additional evidence and can be determined later.

Key insight. A three-month unemployment average of 4.11% is 0.51 point above a prior low of 3.60%. Track which low remains in the lookback window before reading the next release.

Build the comparison in the right order

Compute a rolling three-month mean for each eligible month. Locate the lowest eligible three-month mean in the prior twelve months and subtract it from the current mean. Maintain adequate history at the start of the sheet. When the previous low expires, the measured gap can fall even though the latest unemployment rate has barely changed.

Why two reproductions can disagree

A reproduction from the unemployment data available today can use revised or differently rounded inputs. The real-time indicator preserves information associated with the releases used to calculate it. Record whether a number came directly from the published real-time series or from a reconstruction. A threshold in today’s data cannot retroactively create a trading signal on an earlier news release.

Walk a low out of the twelve-month window

Write thirteen three-month averages across a sheet, one per month, and circle the minimum among the first twelve. The thirteenth is the current average. Move forward one month by dropping the oldest comparison value and adding the previous current value to the history. Only after updating the comparison set should you subtract its minimum from the new current average.

Suppose the oldest eligible low is 3.60% and the next-lowest value is 3.78%. A current mean of 4.11% gives a gap of 0.51 point against the lower benchmark. If the 3.60% observation expires and the new current average is 4.15%, the gap relative to 3.78% is 0.37 point. The latest average rose while the rule’s signal weakened. A reader following only the latest unemployment print misses the moving benchmark.

Around the boundary, retain the available precision and check the published indicator. Rounding three monthly rates, averaging them and rounding again can place a reconstruction on the other side of a displayed threshold. Do not treat a few hundredths of disagreement as evidence of a different business cycle before checking dates and vintage.

An indicator intended as an alarm can be valuable even though it does not supply an end date, speed of contraction or currency direction. List the decision the alarm would prompt, such as examining a broader set of labor and spending data. To turn it into a trading rule you would need a separate, tested link from publicly available values to realizable future returns, with clear treatment of false alarms.

What ten years of US releases show

For each month compute the mean of its US unemployment rate and the prior two months. Compare this mean with the minimum of twelve preceding three-month means. The exercise uses the unemployment history in the retrieved file, including later revisions.

Use the rows below as a check on the method. The first complete observation is 2016-01 and the last is 2025-09. There are 117 usable monthly observations in the 2016 through 2025 study window. A row needs both inputs and the second calculation described below. A missing source observation is left out if the calculation depends on it. An empty cell has a different meaning from a genuine zero. The table shows selected rows and retains full precision for the calculations behind the displayed values.

Start with 2016-03, the observation closest to the sample median for reconstructed sahm gap in percentage points. Enter 4.900 for current three-month unemployment average and 4.900 for lowest prior twelve-month three-month average. Applying the rule gives 0.000. Write the unit beside the result before rounding. It determines whether this number belongs in a dollar ledger, a rate comparison or a column measuring a change. Writing an extra percent sign or moving the decimal two places can produce a neat calculation that answers an entirely different question.

Selected complete periods, ordered by date. The low, middle and high refer to the calculated result.
PeriodCurrent three-month unemployment averageLowest prior twelve-month three-month averageReconstructed Sahm gap in percentage points
2016-014.9675.033-0.067
2016-034.9004.9000.000
2020-0613.0003.5679.433
2021-095.0675.467-0.400
2025-094.3334.1000.233

For a sense of ordinary variation, the middle half of the observations lies between -0.067 and 0.267 on the reconstructed sahm gap in percentage points measure. The endpoints of the full sample are -0.400 in 2021-09 and 9.433 in 2020-06. These are different comparisons. A reader budgeting for a single decision might want to test an adverse endpoint. A reader describing a typical period has more reason to look at the middle half. Neither range gives a promised boundary for what a later observation can be.

What if an older low rolls out while unemployment rises?

Increase the current three-month mean by a tenth of a point and the minimum in the preceding window by 0.17 point. This describes a possible change between windows, not an adjustment to a single worker record. Recalculate the gap and its distance from 0.50.

Scenario arithmetic based on 2016-03. Only the baseline draws its two inputs from the historical worksheet.
IllustrationCurrent three-month unemployment averageLowest prior twelve-month three-month averageReconstructed Sahm gap in percentage points
Baseline4.9004.9000.000
First change5.0004.9000.100
Second change4.9005.070-0.170
Both changes5.0005.070-0.070

Read the first two rows across. The first alteration is 0.100 in the units described as points added to the current average. With the other input fixed at 4.900, the result moves from 0.000 to 0.100, a signed change of 0.100. Now return to the original inputs before applying the second alteration of 0.170 in the units described as points added to the prior low. That separate result is -0.170. Starting the second trial from the already altered first row would accidentally test both changes at once.

With both changes in place the answer becomes -0.070, or -0.070 relative to the baseline. Adding the two isolated changes gives -0.070. The difference between the joint and additive calculations is 0.000 at the displayed precision. The two increments add exactly in this calculation. That makes it possible to attribute their combined effect without choosing which came first.

A higher current mean can coexist with a smaller signal when a sufficiently low earlier observation expires. You need the dates and values of the old and new minima to explain that combination.

A real-time signal requires the rolling benchmark known then. A revised unemployment history can move the minimum to another month.

Reconstruct the largest move between adjacent releases

Between 2020-03 and 2020-04, reconstructed sahm gap in percentage points went from 0.267 to 4.000. Its signed change of 3.733 had the greatest absolute size among the 116 complete adjacent month pairs. To find it, subtract each earlier value from its neighbor, take absolute values solely to rank the pairs and restore the sign when reporting the winner. Rank changes rather than levels. The highest level and the fastest move need not occur at the same time.

The first input went from 3.833 to 7.567. The second went from 3.567 to 3.567. Replace just the first earlier input with its later value. This gives 4.000, which accounts for 3.733 of the observed change on that replacement order. Replacing the second input then accounts for 0.000. The two amounts reconcile to 3.733. This arithmetic answers which measured input changed the answer. Evidence outside the worksheet is needed to explain why the input itself moved.

Try replacing the second input first. Its intermediate result is 0.267, followed by the same final 4.000. The two orders are useful when discussing current three-month unemployment average alongside lowest prior twelve-month three-month average. If the assigned contributions depend on order, show both or give the order explicitly. An unqualified statement that one factor accounted for the entire change might depend on an unnoticed choice about which input the calculation held constant.

Check a less dramatic interval

For comparison, the median absolute adjacent change was 0.033. The pair nearest that magnitude ran from 2019-10 to 2019-11, moving from -0.033 to 0.000. For that pair the first input changed by 0.000 and the second by -0.033. Put these inputs alongside the largest-move pair. A single dramatic date is a poor place to learn the scale of an ordinary observation, especially if the date was chosen after reviewing the entire history.

Difference from the published real-time series in basis points

Subtract the published real-time indicator for the same month from the reconstructed gap and multiply by one hundred. A small discrepancy may reflect rounding. Differences across vintages warrant checking the original release.

Keep the second result on a separate line in the working file. In 2024-08, the first measure, reconstructed sahm gap in percentage points, was 0.567. The second, difference from the published real-time series in basis points, was -0.333. This date was selected because the second result lies closest to its own sample median, rather than because it gives the most favorable comparison. Start from the inputs in the table if you reproduce it. A second measure can have a different scale, denominator and sign interpretation from the first.

A separate calculation using the same records and the assumptions stated above.
PeriodCurrent three-month unemployment averageLowest prior twelve-month three-month averageDifference from the published real-time series in basis points
2016-054.9674.9009.667
2020-0711.4673.567-10.000
2024-084.1673.600-0.333
2025-094.3334.1000.333

At the lower end of this calculation, 2020-07 gives -10.000 with inputs of 11.467 and 3.567. At the upper end, 2016-05 gives 9.667 with inputs of 4.967 and 4.900. Check the formula before interpreting the order. A dollar outlay, a net receipt, a growth rate and a distance from a benchmark answer different economic questions even when their values can all be ranked from low to high. The real-time and revised histories answer different questions. The fifty-basis-point threshold belongs to the gap, not to its reconstruction error.

Reading the final observation without hindsight

Cover the last row and write two estimates for reconstructed sahm gap in percentage points. One uses the latest earlier observation, 0.133 from 2025-08. The other uses the median of the preceding 116 records, 0.000. Reveal 2025-09. The observed result is 0.233, so actual minus estimate is 0.100 for the recent observation and 0.233 for the historical middle. Keep both errors with the original estimates. Quietly replacing a poor baseline with a better one after the reveal removes the point of the exercise.

Next inspect the final record itself. It contains 4.333 for current three-month unemployment average and 4.100 for lowest prior twelve-month three-month average, producing 0.333 on the second decision measure. Someone reproducing this result from rounded figures might differ in the last decimal. A large difference demands a different explanation. Check the observation period, quote direction or growth interval and whether the same quantity appears twice. Observation, release vintage and dates in the minimum window provide a starting point for an audit. Write down any imposed assumption that cannot be found in those records.

Did the indicator separate stronger and weaker US dollar periods?

To connect the lesson with the US currency market, compare the worksheet with a different series. Use the Federal Reserve nominal broad dollar index, measured at the last available observation in each period. A rise means the dollar strengthened against its weighted basket. Measure the percentage change from one period-end to the next. That change is a basket return. A specific bilateral currency, a card settlement or a leveraged EUR/USD position can give a different sign and amount.

Take the first 58 matched transitions as a calibration segment. Their starting-period median for reconstructed sahm gap in percentage points was 0.000. Freeze that cutoff. In the remaining 58 transitions, group each starting observation by whether it was at least that value. Measure the subsequent broad dollar return within each group. Keeping adjacent pairs only prevents a missing month or quarter from silently becoming a longer observation with a different holding period. Missing values never count as zero returns.

Later sample only. Group membership uses the worksheet at the start of each adjacent interval.
Later-period groupRecordsMean USD move %Median USD move %USD rose
At or above 0.00040-0.119-0.30916
Below 0.000180.6360.31612

The higher group contains 40 later observations with a mean next-period dollar move of -0.119%. The lower group contains 18 with a mean of 0.636%. Their difference is -0.755 percentage points on a high-minus-low convention. A gap is descriptive evidence for this sample and definition. Before attributing it to this measure, check whether the groups cluster in different years and whether the median tells the same story as the mean.

Release timing limits what the calculation can establish. An observation labeled with a month or quarter need not have been public at its last market close. Historical series can also change after revision. For a prospective test, replace each worksheet label with the timestamp when both inputs could actually have been known, then start the currency return after that time. Preserve a release vintage and executable prices. The table here is a retrospective comparison and supplies no entry price, success probability or trading performance record.

Would the real-time and reconstructed indicators agree at the trigger?

Read the real-time Sahm indicator directly for the selected month. Compare it with the reconstruction from the retrieved unemployment history. Check whether the two values fall on the same side of the 0.50-point trigger. Keep a numeric discrepancy separate from a disagreement in trigger status.

A researcher checks the vintage if the absolute difference exceeds 0.05 percentage point, even where the trigger classification agrees.

Decision check. Differences are in percentage points. The comparison allowance is 0.050 percentage points.
ObservationPublished real-time gap in pointsReconstructed unemployment gap in pointsSecond minus first
2016-01-0.030-0.067-0.037
2020-078.0007.900-0.100
2025-060.1700.2330.063
2025-090.2300.2330.003

Use the last complete case, 2025-09, as a reconciliation. The first calculation gives 0.230 and the second gives 0.233. Subtract the first from the second to obtain 0.003 percentage points. A positive result has this particular interpretation: the revised-data reconstruction gives a larger gap than the published real-time series. The absolute distance is 0.003 against the assumed allowance of 0.050. Restoring the sign after checking the size of the difference matters when one party pays an amount that another party receives.

For 2025-06, the absolute difference was 0.063 percentage points. This is the observation nearest the stated allowance, with a distance from that boundary of 0.013. The underlying values were 0.170 and 0.233. Check a borderline case at full precision before assigning it to either side. If you change the allowance, carry the new rule through all dates. Applying it to just the most interesting row would quietly change the selection method.

How often did the difference matter under these assumptions?

The absolute difference exceeded 0.050 percentage points in 29 of 117 complete records. It stayed at or below that level in 88. The largest observed absolute difference was 0.100 in 2020-07, against a middle observation by magnitude of 0.033. This count gives each observation equal weight. A person whose largest obligations or research decisions cluster in a small number of months can experience a very different total effect from the fraction of dates shown here.

A revision error near the trigger can change a classification. A larger discrepancy far from the boundary may leave both readings on the same side.

Using 0.50 percentage point as the trigger in each series, the two classifications differed in 0 of the 117 retained observations. No retained observation landed on different sides of the threshold. A change in the trigger classification is separate from the absolute discrepancy check above.

Build the calculation in two spreadsheet cells

Put current three-month unemployment average in column A and lowest prior twelve-month three-month average in column B. Use row one for labels and enter the two inputs from 2016-03 in row two. The numeric entries are 4.90000 and 4.90000 when displayed to five decimals. In this worksheet, a percentage quoted as 4.25 is entered as the number 4.25, not as the spreadsheet value 0.0425. For dollar or currency amounts, enter numbers and use cell formatting for currency symbols so the program can still do arithmetic.

In C2 enter =A2-B2. The expected result for reconstructed sahm gap in percentage points is 0.00000 with unrounded inputs. Repeat the formula on the next dated row before copying it through a longer file. A second manual check on a different observation is more informative than assuming that a formula which worked on a single coincidence will work everywhere.

Reporting a percentage increase in unemployment instead of a point change

The threshold of 0.50 applies to a percentage-point difference between averages. A relative percentage increase has another denominator and cannot be compared with that threshold. Keep percentage-point labels beside the calculated value, especially when a charting platform offers percent change by default.

To diagnose that problem, enter 100*(A2/B2-1) in a separate comparison cell. With these inputs it displays 0.00000. Keep the column labeled as a diagnostic. A superficially similar number can have a different unit, denominator or sign. Where two formulas give the same number on a particular row, that agreement is a reason to test a second case with deliberately different values. It does not establish that the formulas are interchangeable.

Use 4.10% for the current mean and 3.60% for the prior low. The difference should land exactly at the half-point boundary. Then move the prior low to 3.70% and verify a reading of 0.40 point while the current mean stays fixed.

How to identify the price setup on a EUR/USD chart

For EUR/USD, the economic question is what was unexpected at the moment of publication. An already anticipated crossing may coincide with a small move, and a non-crossing can accompany a large surprise. Give the simulated order its own price-based conditions.

Use a five-minute demo chart with a named time zone, such as America/New_York. The first decision is where an existing swing could attract orders. In the drawing the prior high sits at 1.09337. Mark it while the chart still ends before the next challenge. The subsequent wick reaches 1.09419, above the earlier reference, and the market returns through it. ICT describes that sequence as a liquidity sweep. Candles show traded or quoted prices on a feed, so the sequence does not establish any individual institution’s identity or the number of orders at the swing.

The next test is displacement through the earlier internal swing at 1.09131. In this schematic the outer wicks of bars five and seven leave a three-candle fair value gap between 1.09146 and 1.09196. Write down both boundaries before testing the retest. A wick across the initial level without displacement is a different observation. A displacement without a later return can leave a correctly identified setup unfilled. Preserve each outcome so a replay does not silently select only the attractive-looking retracements.

Chart 1. Identify the sweep and three-candle imbalance
EUR/USD | hypothetical spot pricesPrior high 1.09337Sweep 1.09419Prior swing 1.09131Gap 1.09196 / 1.09146Green: up Red: down Blue: reference Purple: sweep Gold: gap
Schematic EUR/USD bars. Compare the wick of bar five with the wick of bar seven. The middle bar provides displacement. The shaded interval records the gap between the outside wicks.

How to trade the retest on demo

For the hypothetical short position, the order waits at 1.09172 after the gap has formed. The protective stop is 1.09466, beyond the swept extreme. The target is 1.08643, which requires the quoted pair to move lower. Cancel the order if price reaches invalidation before entry, the agreed observation window expires, or the executable spread makes the original risk budget infeasible. A chart mid touching a line does not prove the customer’s bid or offer filled. In a live record those two events need separate timestamps.

The second drawing shows one favorable branch solely to establish the accounting. A losing branch exits at the stop under an exact-fill assumption. A third branch never enters and incurs no modeled round-trip charge. If stop and target both lie within one observed bar, the wick alone does not reveal their order. Seek finer time-stamped evidence or retain an ambiguous result rather than labeling the trade a winner. For an event study, select the exit, expiry and treatment of ambiguous sequences before viewing the release-day chart.

Check the original publication time in Eastern time against the chart clock. A price move before the release cannot be attributed to the reported value merely because both appear under the same calendar date. A trade record needs the pre-release spread, the available quote after the news and the point at which the order became eligible.

Chart 2. Price the entry, protective stop and target
EUR/USD | hypothetical spot pricesEntry 1.09172Stop 1.09466Target 1.08643Green: up Red: down Blue: reference Purple: target
Illustrative continuation after a retest. Lines show the planned short order. The favorable path is one possible sequence and carries no probability estimate.

Worked trade in a dollar-denominated demo account

Illustrative position. The prices are separate from the historical observations and the outcome has no assigned probability.
FieldHypothetical execution
Pair and sideEUR/USD short
Size21,000 EUR
Entry1.09172
Protective stop1.09466
Target1.08643
Gross risk and potential reward$61.74 and $111.09
Assumed round-trip charges$3.21

First subtract the entry and stop in absolute price units. The distance is 0.00294 USD per EUR, compared with 0.00529 to the target. Multiply each by 21,000 base units. That yields 61.7400 USD of gross risk and 111.0900 USD of potential gross gain. The figures are already in the USD account currency. Keep the full numerical precision through this step.

Gross planned risk is $61.74 and potential gain is $111.09. Divide gain by risk: $111.09 / $61.74 = 1.80. On a risk-to-reward convention that is 1:1.80. Now assume $3.21 in total round-trip charges in either completed branch, with executable entry and exit prices in the example. Net risk is $61.74 plus $3.21, or $64.95. Net potential reward is $111.09 minus $3.21, or $107.88. The net R:R is 1:1.66 from $107.88 / $64.95. These ratios use full-precision amounts before display rounding.

If the stop slipped one further chart increment of 0.00010, the modeled exit would move to 1.09476 and the completed loss including the assumed charge would become $67.05. Compare this with the original $64.95. Slippage, funding and market access can all affect an actual instrument. The purpose of this branch is to make the size and account-currency conversion explicit. A promising reward-to-risk ratio says nothing about the frequency with which this particular entry, protective exit and target are actually executable.

For the chart terminology and session conventions used in a replay, see fair value gaps, order blocks, the London session, market structure shifts, the Judas swing.

Sources and data conventions

St. Louis Fed real-time Sahm indicator. Historical series used in the calculations: UNRATE observations and units, SAHMREALTIME observations and units, DTWEXBGS observations and units.

Frequently asked questions

What is the Sahm Rule threshold?

A gap of at least half a percentage point between the current three-month unemployment average and its preceding twelve-month low meets the rule. This is measured in percentage points, not relative percent growth.

Does one bad unemployment print trigger the rule?

It can contribute, but the comparison uses smoothed three-month averages and depends on the earlier minimum. Recalculate the entire window instead of comparing only the two latest monthly rates.

Does a trigger officially declare a US recession?

It is an indicator constructed from labor-market data. The US business-cycle chronology uses broader evidence. A signal can be informative without settling the official dating of a recession.

Can the gap be negative?

With the minimum taken over preceding months, a new lower average can produce a negative reading. Do not force the output to zero unless your explicitly stated alternative definition includes the current month.

Which series belongs in a release-day FX backtest?

Use information available at the decision time, including the corresponding vintage and publication timestamp. A later reconstruction can help explain revisions but should not replace what was available to the market that day.

For related US dollar examples, read U-3 and U-6 unemployment, US job openings, ADP and payroll employment, and interpreting a news release.